LightMamba introduces a novel approach to accelerate Mamba models on FPGA hardware through quantization and architecture co-design, achieving significant performance improvements.

• Delivers linear computational complexity with sequence length compared to transformer-based LLMs
• Addresses scattered activation outliers and complex computation dependencies that hinder efficient acceleration
• Implements specialized hardware optimization techniques including computation reordering, tiling, and fusion
• Demonstrates how hardware-software co-design can overcome implementation challenges for state space models

This research is particularly valuable for engineering teams developing hardware acceleration for AI models, providing a pathway to deploy efficient Mamba models in resource-constrained environments.

LightMamba: Efficient Mamba Acceleration on FPGA with Quantization and Hardware Co-design